OpenAI has introduced its latest series of artificial intelligence models, GPT-5.6, to a limited group of users approved by the U.S. government, amid concerns over the potential risks posed by highly advanced AI systems in cybersecurity. The announcement, made yesterday by the San Francisco-based company, includes three models named Sol, Terra, and Luna, with Sol positioned as the flagship model equipped with the most advanced cybersecurity features.

This restricted rollout follows a recent move by Anthropic, a competitor, which was compelled to remove its model Fable from public access due to an export-control directive issued by the U.S. administration. Fable, a safeguarded version of Anthropic’s earlier Mythos model, which also boasted significant cybersecurity capabilities, was withdrawn after officials identified a security flaw.

OpenAI has been working closely with government agencies to preview GPT-5.6 before its release. The administration requested a controlled distribution, leading to access being granted only to a small group of government-approved customers. The company indicated plans to expand availability gradually in the coming weeks but did not disclose the identities of those granted access.

While OpenAI accepts this limited release as a temporary measure, it expressed concerns over such government-controlled access becoming a permanent practice. The company cautioned that restricting the tools could hinder developers, enterprises, cybersecurity professionals, and international partners who rely on these technologies.

This development comes as the U.S. government, under President Donald Trump’s recent executive order, is moving to establish a voluntary regulatory framework aimed at ensuring new AI models are safe before deployment, particularly those with cybersecurity applications.

The release of Mythos by Anthropic in April heightened global unease regarding AI’s dual-use potential—its ability to identify and exploit software vulnerabilities at an unprecedented scale, including those affecting critical infrastructure. This has raised alarms among governments and industry leaders about the misuse of such technologies by malicious actors or foreign adversaries.

OpenAI emphasized that GPT-5.6 Sol incorporates enhanced safeguards to mitigate risks related to high-risk activities, sensitive cyber requests, and recurrent misuse. The company stated that the model is designed to assist users in identifying and remedying vulnerabilities rather than facilitating complete cyber attacks. It noted that significant resources were dedicated to safety testing, including layered protections within the model’s training, real-time response monitoring, account vetting, and attempts to prevent “jailbreaks”—techniques users might employ to bypass security controls.

Besides cybersecurity, OpenAI noted that the GPT-5.6 series advances capabilities in scientific research, programming, and goal-oriented tasks. The company highlighted its commitment to ensuring the technology benefits defenders working to fortify digital systems while continuing to collaborate with the government on establishing formal protocols for future AI model releases.